IEEE Consumer Electronics Magazine - May/June 2023 - 67

FIGURE 4. Blockchain simulation result for various cases. (a) Blockchain simulation result for Case I.
(b) Blockchain simulation result for Case II.
1) performance of ML model under data poisoning
attacks and 2) performance of ML model without
data poisoning attacks. In the first case, we show
how the data poisoning attacks can reflect the ML
model when the data are on some cloud storage
(i.e., not in the blockchains), whereas in the second
case, we show the effect when the data are in
the blockchains (i.e., without data poisoning
attacks).
Performance Metrics
To depict the correct performance analysis of
the ML models under data poisoning attack, we
have used four important ML metrics. Assume
that TPn, TNn, FPn, and FNn denote the " true
positive, " " true negative, " " false positive, " and
" false negative, " respectively, for the nth experiment.
The following metrics are then defined.
Accuracy Accuracy score is a very important
performance metric, which can be used to depict
the correct prediction score of the ML models.
The accuracy can be defined mathematically as
Accuracy ¼
TPnþTNn
TPnþTNnþFPnþFNn
.
Recall Recall is one of the most useful performance
metrics that can be used in ML to show
the true positive rate of the model. It can be
defined as Recall ¼
TPn
TPnþFNn
.
Precision Precision is another very important
performance metric, which depicts the precision
of the ML model. It used to compare the number
of true positives that a model claim with the
total true positives in the dataset. It is defined as
Precision ¼
TPn
TPnþFPn
.
May/June 2023
Salt noise injecting attack: In the case of the
salt noise injecting attack, an attacker gains
the access of the data storage and then uses
the standard salt noise function to corrupt the
data in a meaningful way.
Label flipping attack: In the case of the label
flipping attack, the attacker first gains the
67
F1 Score F1 score is most important metric
used in ML to show how the ML model trades
with both recall and precision. It is the weighted
average of recall and precision of the ML model.
It is defined as F1 Score ¼
2TPn
2TPnþFPnþFNn
.
Performance of ML Model Under Data
Poisoning Attacks
This section describes about the performance
of the ML model under data poisoning
attacks.
Correct data are crucial for the ML model to
perform better. If the training data are corrupted
by any means, they can hamper the overall performance
of the ML model. Data that are present
in any centralized database can easily be corrupted
as the database does not ensure the
tamper proof ability of the data storage. In contrast,
if the data are present in any blockchain
system, they are guaranteed to be tamper proof
by the design of the blockchain system. So, if the
data are not stored in any blockchain system,
various data poisoning attacks are possible on
the training data of the ML model. In the following,
we discuss about two types of data poisoning
attacks, namely 1) salt noise injecting attack
and 2) label flipping attack.

IEEE Consumer Electronics Magazine - May/June 2023

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - May/June 2023